A Remote Sensing-Based Daily Stream Water Temperature Model for Gridded, High-Resolution Predictions at Subcontinental Scales

Stream water temperature (SWT) is fundamental to studies in water quality and ecology, with effects ranging from drinking water treatment chemistry to lotic species metabolism.  Assessment of current and future SWT at large scales requires data at high spatial and temporal resolution, which can be supported by modeled datasets at considerably higher spatial resolution and extent than is feasible with monitoring networks.  While several models support moderate- to high-resolution SWT estimation and prediction for unmonitored streams over large domains and global SWT modeling has been conducted at 10 km/daily resolution, no high-resolution (1 km/daily) dataset exists for the contiguous United States (CONUS) or other subcontinental domains. In addition, current high-resolution models are not optimized for gridded processing over large regions and are thus computationally impractical for high-density model runs over subcontinental domains.  We address this limitation by enhancing an existing remote-sensing, ungauged, high-resolution SWT model, TempEst 2 (“temperature estimation, version 2”), to support computationally efficient analyses, including data retrieval and processing, over large blocks of input pixels. TempEst 2 is particularly suited to this optimization because the model only uses data near the point of interest, allowing a direct mapping from input to output grids without the need to process entire watersheds. Using the optimized model TempEst 2-FAST (“fast analysis in space and time”), we present progress on a 1 km/daily resolution SWT dataset over the CONUS (~8 million km²).

TempEst 2 has a median CONUS validation RMSE of 2.0 C, NSE of 0.91, and bias of 0.10%, within the typical performance range of regional to global ungauged daily SWT models (RMSE ~ 1.8-3.2 C).  While TempEst 2 is trained on the United States Geological Survey SWT gauge network, it uses globally-available satellite-based or gridded data (e.g., land surface temperature, humidity) for prediction, supporting straightforward application outside the CONUS given a suitable local gauge network for training and validation. TempEst 2 is also relatively robust to spatial gaps in gauge network coverage and to overall sparse gauge networks, maintaining reasonable accuracy (approximately a 20% performance penalty) with just 100 gauges across the CONUS (~13 per million km²). Within the CONUS, the model shows consistent performance across a range of geographic and climate conditions, though there is some performance penalty in extrapolating to high-elevation (> 3000 m) sites (a small proportion of streams).  Building on that robust performance and efficient data generation, the CONUS-wide gridded dataset we are developing provides readily-available data for large-domain analyses at far higher resolution than previously possible, with millions of prediction points over ~9,000 days (2001-2024). The availability of high-resolution SWT estimates over the CONUS enables rapid assessment of stream thermal conditions that would otherwise require extensive local fieldwork or modeling efforts. We anticipate that the dataset could be particularly useful for detailed assessments of ecological conditions or regulatory compliance over large regions.